---
license: cc-by-4.0
language:
  - de
  - en
task_categories:
  - text-classification
tags:
  - content-moderation
  - toxicity
  - benchmark
pretty_name: ChatMod Bench
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: test.jsonl
---

# ChatMod Bench

600 German and English chat messages, half toxic. One question per message: would a
moderator in a normal online community remove it or warn the author?

- `chat` (300): gaming and community chat, including harmless look-alikes such as game
  banter, quotes, reports to moderators and negation
- `hatecheck` (300): a sample of HateCheck and German HateCheck

```json
{"id": "cmb-chat-de-1a2b3c4d5e", "text": "...", "label": 1, "lang": "de", "source": "chat", "category": "implied"}
```

## Results

| Model | Accuracy | Macro-F1 |
|---|---:|---:|
| OpenAI omni-moderation-latest | 75.7 % | 0.754 |
| axo-native-small-preview | 75.0 % | 0.746 |
| Qwen3Guard-Gen-8B | 70.2 % | 0.684 |
| Qwen3Guard-Gen-4B | 69.5 % | 0.678 |
| Qwen3Guard-Gen-0.6B | 66.2 % | 0.644 |

Default decision rule of each model. Full numbers in `results.json`.

## Run

```bash
pip install -r requirements.txt
python run_benchmark.py --model omni                  # or axo-native-small-preview, qwen3guard-0.6b/4b/8b
python run_benchmark.py --predictions my_model.jsonl  # {"id": ..., "flagged": true, "score": 0.93}
```

## License

CC BY 4.0. The `hatecheck` part is from HateCheck (Röttger et al., ACL 2021) and
Multilingual HateCheck (Röttger et al., WOAH 2022), both CC BY 4.0. Contains offensive
language by design.